EQUIPA memory profile

Type: agentic-system-analyses/types/agentic-system-memory-profile.md

Comparison rationale

This profile classifies the supplied accepted records, not source files or an incumbent profile. Every axis is partial because the recorded inventory retains material alternatives with missing admission, provenance, transformation or later-consumer facts. A wired finding is implementation-level capacity. It does not establish enabled features, exercised model reliance, semantic faithfulness or benefit.

The reconciliation supersedes the epistemic episode, lesson and Q-value object declarations with MEM-OBJ-agent-episodes, MEM-OBJ-lessons-learned and MEM-OBJ-q-value. This profile uses those canonical objects; the independently scoped epistemic routes and effectiveness-score part remain valid. No route/container supersession is inferred from overlap.

The retained MCP disagreement concerns implementation status, not query syntax. MEM-RTE-memory-queries warrants an afforded pull route; EPI-RTE-mcp-retrieval calls its inspected handler wired but does not establish subsequent client incorporation. The profile retains the capability-level pull witness and leaves downstream authority incomplete rather than upgrading it from that handler alone.

SQLite holds text, symbolic scores, dense vectors and graph-edge rows. Those forms do not imply four storage engines. Dense vector similarity supports parametric representation and inferred-embedding selection, not model-weight learning. The configured embedding model and historical vector compatibility remain unresolved in MEM-OBJ-embedding. Numeric Q/effectiveness and graph IDs remain symbolic at their fixed-rule consumers.

Historical text has knowledge authority at task inference; changed role text has instruction authority. Retained experience has learning authority at the maintenance/derivation consumer, while access fields have ranking, routing or enforced eligibility effects at host code. Existing-rule duplicate criteria and version-performance criteria have validation force at their own admission consumers. These forces are local: MEM-RTE-lesson-maintenance can withdraw generic injection without revoking the MCP read surface. The source-label and graph-ID qualifications in MEM-RTE-experience-retrieval prevent idealized coherence or usefulness claims, not the recorded positive selectors.

Consolidation is established by deterministic continuation extraction and bounded reconstruction in MEM-RTE-context-recovery. SIMBA's requested synthesis and admitted rule format do not establish semantic novelty. EPI-RTE-simba-rules keeps the content relation indeterminate without actual rule instances. No synthesize value is invented to match its name. Positive Q/frequency changes raise salience; negative changes lower it. Pruning and GEPA reversion retain history while withdrawing reliance. These curation findings do not establish causal improvement.

Automatic write agency is assessed at each admission, not at workflow launch or physical I/O. Known automatic reflection replacement, session reconstruction and initiative appends coexist with unresolved original state-file/plan/table control. Interactive human supplying/editing is only afforded manual admission. ForgeSmith's default manual patch gate does not govern GEPA or autoresearch and does not by itself establish that a proposal was manually installed.

Push signals name actual selectors: task/role/project/initiative matches, coarse state/expiry/budget limits, lexical overlaps and optional dense similarity. An explicit MCP query remains pull, including its requested response. The memory records do not establish an inferred-judgment selector merely because a model generated the stored text. Nor does the existence of an output path establish a later reload.

Automatic trace-fed continuation artifacts qualify alongside cross-task rules and prompt changes. Their later consumers are conditional but wired; full conversation restoration is not. MEM-RTE-context-recovery retains the history gate, lexical checkpoint ordering and possible simultaneous session/checkpoint inclusion. Recorded session tool metadata supports tool-traces; streaming turn/compaction signals support event-streams only on that sampled runtime-state part. An agent narrative or JSON wrapper is not relabeled as an event stream. Neither training-registry metadata nor inference service use establishes weight training, and no bounded complete absence permits a system-wide trace-learning no.

Source-read log: none. No target source was read to supplement the accepted record set. Missing source facts remain limitations rather than new evidence or record declarations.